A novel neuro-fuzzy method for linguistic feature selection and rule-based classification

This paper proposes a new interpretable neuro-fuzzy classification mechanism. The proposed neuro-fuzzy structure is different from other data analysis mechanisms previously invented in pattern recognition. General mechanisms focus mainly on creating predictive data models whereas some useful information inside the process may be ignored. The proposed mechanism is designed based on the consideration of feature selection and rule extraction. It is a three-layer feedforward network. Its structure can be comprehended to logical rules using only selected important features. We construct a new classification algorithm by using a small number of features that represent an informative subset of a given dataset. This classifier can produce good classification results from the direct calculation or from logical rule extraction. Pleasant performance of classification results are acquired from 10-fold cross validation testing on several standard datasets.